|
| 1 | +// Copyright 2024 Xiaomi Corporation |
| 2 | + |
| 3 | +// This file shows how to use a silero_vad model with a non-streaming Paraformer |
| 4 | +// for speech recognition. |
| 5 | + |
| 6 | +import com.k2fsa.sherpa.onnx.*; |
| 7 | +import java.util.Arrays; |
| 8 | + |
| 9 | +public class VadNonStreamingParaformer { |
| 10 | + public static Vad createVad() { |
| 11 | + // please download ./silero_vad.onnx from |
| 12 | + // https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models |
| 13 | + String model = "./silero_vad.onnx"; |
| 14 | + SileroVadModelConfig sileroVad = |
| 15 | + SileroVadModelConfig.builder() |
| 16 | + .setModel(model) |
| 17 | + .setThreshold(0.5f) |
| 18 | + .setMinSilenceDuration(0.25f) |
| 19 | + .setMinSpeechDuration(0.5f) |
| 20 | + .setWindowSize(512) |
| 21 | + .build(); |
| 22 | + |
| 23 | + VadModelConfig config = |
| 24 | + VadModelConfig.builder() |
| 25 | + .setSileroVadModelConfig(sileroVad) |
| 26 | + .setSampleRate(16000) |
| 27 | + .setNumThreads(1) |
| 28 | + .setDebug(true) |
| 29 | + .setProvider("cpu") |
| 30 | + .build(); |
| 31 | + |
| 32 | + return new Vad(config); |
| 33 | + } |
| 34 | + |
| 35 | + public static OfflineRecognizer createOfflineRecognizer() { |
| 36 | + // please refer to |
| 37 | + // https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-paraformer/paraformer-models.html#csukuangfj-sherpa-onnx-paraformer-zh-2023-03-28-chinese-english |
| 38 | + // to download model files |
| 39 | + String model = "./sherpa-onnx-paraformer-zh-2023-03-28/model.int8.onnx"; |
| 40 | + String tokens = "./sherpa-onnx-paraformer-zh-2023-03-28/tokens.txt"; |
| 41 | + |
| 42 | + String waveFilename = "./sherpa-onnx-paraformer-zh-2023-03-28/test_wavs/3-sichuan.wav"; |
| 43 | + |
| 44 | + WaveReader reader = new WaveReader(waveFilename); |
| 45 | + |
| 46 | + OfflineParaformerModelConfig paraformer = |
| 47 | + OfflineParaformerModelConfig.builder().setModel(model).build(); |
| 48 | + |
| 49 | + OfflineModelConfig modelConfig = |
| 50 | + OfflineModelConfig.builder() |
| 51 | + .setParaformer(paraformer) |
| 52 | + .setTokens(tokens) |
| 53 | + .setNumThreads(1) |
| 54 | + .setDebug(true) |
| 55 | + .build(); |
| 56 | + |
| 57 | + OfflineRecognizerConfig config = |
| 58 | + OfflineRecognizerConfig.builder() |
| 59 | + .setOfflineModelConfig(modelConfig) |
| 60 | + .setDecodingMethod("greedy_search") |
| 61 | + .build(); |
| 62 | + |
| 63 | + return new OfflineRecognizer(config); |
| 64 | + } |
| 65 | + |
| 66 | + public static void main(String[] args) { |
| 67 | + |
| 68 | + Vad vad = createVad(); |
| 69 | + OfflineRecognizer recognizer = createOfflineRecognizer(); |
| 70 | + |
| 71 | + // You can download the test file from |
| 72 | + // https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models |
| 73 | + String testWaveFilename = "./lei-jun-test.wav"; |
| 74 | + WaveReader reader = new WaveReader(testWaveFilename); |
| 75 | + |
| 76 | + int numSamples = reader.getSamples().length; |
| 77 | + int numIter = numSamples / 512; |
| 78 | + |
| 79 | + for (int i = 0; i != numIter; ++i) { |
| 80 | + int start = i * 512; |
| 81 | + int end = start + 512; |
| 82 | + float[] samples = Arrays.copyOfRange(reader.getSamples(), start, end); |
| 83 | + vad.acceptWaveform(samples); |
| 84 | + if (vad.isSpeechDetected()) { |
| 85 | + while (!vad.empty()) { |
| 86 | + SpeechSegment segment = vad.front(); |
| 87 | + float startTime = segment.getStart() / 16000.0f; |
| 88 | + float duration = segment.getSamples().length / 16000.0f; |
| 89 | + |
| 90 | + OfflineStream stream = recognizer.createStream(); |
| 91 | + stream.acceptWaveform(segment.getSamples(), 16000); |
| 92 | + recognizer.decode(stream); |
| 93 | + String text = recognizer.getResult(stream).getText(); |
| 94 | + |
| 95 | + if (!text.isEmpty()) { |
| 96 | + System.out.printf("%.3f--%.3f: %s\n", startTime, startTime + duration, text); |
| 97 | + } |
| 98 | + |
| 99 | + vad.pop(); |
| 100 | + } |
| 101 | + } |
| 102 | + } |
| 103 | + } |
| 104 | +} |
0 commit comments